Subject: rtmutex: Handle the various new futex race conditions
From: Thomas Gleixner <tglx@linutronix.de>
Date: Fri, 10 Jun 2011 11:04:15 +0200

RT opens a few new interesting race conditions in the rtmutex/futex
combo due to futex hash bucket lock being a 'sleeping' spinlock and
therefor not disabling preemption.

Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
---
 kernel/futex.c                  |   77 ++++++++++++++++++++++++++++++++--------
 kernel/locking/rtmutex.c        |   36 +++++++++++++++---
 kernel/locking/rtmutex_common.h |    2 +
 3 files changed, 94 insertions(+), 21 deletions(-)

--- a/kernel/futex.c
+++ b/kernel/futex.c
@@ -1717,6 +1717,16 @@ static int futex_requeue(u32 __user *uad
 				requeue_pi_wake_futex(this, &key2, hb2);
 				drop_count++;
 				continue;
+			} else if (ret == -EAGAIN) {
+				/*
+				 * Waiter was woken by timeout or
+				 * signal and has set pi_blocked_on to
+				 * PI_WAKEUP_INPROGRESS before we
+				 * tried to enqueue it on the rtmutex.
+				 */
+				this->pi_state = NULL;
+				free_pi_state(pi_state);
+				continue;
 			} else if (ret) {
 				/* -EDEADLK */
 				this->pi_state = NULL;
@@ -2576,7 +2586,7 @@ static int futex_wait_requeue_pi(u32 __u
 	struct hrtimer_sleeper timeout, *to = NULL;
 	struct rt_mutex_waiter rt_waiter;
 	struct rt_mutex *pi_mutex = NULL;
-	struct futex_hash_bucket *hb;
+	struct futex_hash_bucket *hb, *hb2;
 	union futex_key key2 = FUTEX_KEY_INIT;
 	struct futex_q q = futex_q_init;
 	int res, ret;
@@ -2635,20 +2645,55 @@ static int futex_wait_requeue_pi(u32 __u
 	/* Queue the futex_q, drop the hb lock, wait for wakeup. */
 	futex_wait_queue_me(hb, &q, to);
 
-	spin_lock(&hb->lock);
-	ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to);
-	spin_unlock(&hb->lock);
-	if (ret)
-		goto out_put_keys;
+	/*
+	 * On RT we must avoid races with requeue and trying to block
+	 * on two mutexes (hb->lock and uaddr2's rtmutex) by
+	 * serializing access to pi_blocked_on with pi_lock.
+	 */
+	raw_spin_lock_irq(&current->pi_lock);
+	if (current->pi_blocked_on) {
+		/*
+		 * We have been requeued or are in the process of
+		 * being requeued.
+		 */
+		raw_spin_unlock_irq(&current->pi_lock);
+	} else {
+		/*
+		 * Setting pi_blocked_on to PI_WAKEUP_INPROGRESS
+		 * prevents a concurrent requeue from moving us to the
+		 * uaddr2 rtmutex. After that we can safely acquire
+		 * (and possibly block on) hb->lock.
+		 */
+		current->pi_blocked_on = PI_WAKEUP_INPROGRESS;
+		raw_spin_unlock_irq(&current->pi_lock);
+
+		spin_lock(&hb->lock);
+
+		/*
+		 * Clean up pi_blocked_on. We might leak it otherwise
+		 * when we succeeded with the hb->lock in the fast
+		 * path.
+		 */
+		raw_spin_lock_irq(&current->pi_lock);
+		current->pi_blocked_on = NULL;
+		raw_spin_unlock_irq(&current->pi_lock);
+
+		ret = handle_early_requeue_pi_wakeup(hb, &q, &key2, to);
+		spin_unlock(&hb->lock);
+		if (ret)
+			goto out_put_keys;
+	}
 
 	/*
-	 * In order for us to be here, we know our q.key == key2, and since
-	 * we took the hb->lock above, we also know that futex_requeue() has
-	 * completed and we no longer have to concern ourselves with a wakeup
-	 * race with the atomic proxy lock acquisition by the requeue code. The
-	 * futex_requeue dropped our key1 reference and incremented our key2
-	 * reference count.
+	 * In order to be here, we have either been requeued, are in
+	 * the process of being requeued, or requeue successfully
+	 * acquired uaddr2 on our behalf.  If pi_blocked_on was
+	 * non-null above, we may be racing with a requeue.  Do not
+	 * rely on q->lock_ptr to be hb2->lock until after blocking on
+	 * hb->lock or hb2->lock. The futex_requeue dropped our key1
+	 * reference and incremented our key2 reference count.
 	 */
+	hb2 = hash_futex(&key2);
 
 	/* Check if the requeue code acquired the second futex for us. */
 	if (!q.rt_waiter) {
@@ -2657,9 +2702,10 @@ static int futex_wait_requeue_pi(u32 __u
 		 * did a lock-steal - fix up the PI-state in that case.
 		 */
 		if (q.pi_state && (q.pi_state->owner != current)) {
-			spin_lock(q.lock_ptr);
+			spin_lock(&hb2->lock);
+			BUG_ON(&hb2->lock != q.lock_ptr);
 			ret = fixup_pi_state_owner(uaddr2, &q, current);
-			spin_unlock(q.lock_ptr);
+			spin_unlock(&hb2->lock);
 		}
 	} else {
 		/*
@@ -2672,7 +2718,8 @@ static int futex_wait_requeue_pi(u32 __u
 		ret = rt_mutex_finish_proxy_lock(pi_mutex, to, &rt_waiter);
 		debug_rt_mutex_free_waiter(&rt_waiter);
 
-		spin_lock(q.lock_ptr);
+		spin_lock(&hb2->lock);
+		BUG_ON(&hb2->lock != q.lock_ptr);
 		/*
 		 * Fixup the pi_state owner and possibly acquire the lock if we
 		 * haven't already.
--- a/kernel/locking/rtmutex.c
+++ b/kernel/locking/rtmutex.c
@@ -69,6 +69,11 @@ static void fixup_rt_mutex_waiters(struc
 		clear_rt_mutex_waiters(lock);
 }
 
+static int rt_mutex_real_waiter(struct rt_mutex_waiter *waiter)
+{
+	return waiter && waiter != PI_WAKEUP_INPROGRESS;
+}
+
 /*
  * We can speed up the acquire/release, if the architecture
  * supports cmpxchg and if there's no debugging state to be set up
@@ -342,7 +347,8 @@ int max_lock_depth = 1024;
 
 static inline struct rt_mutex *task_blocked_on_lock(struct task_struct *p)
 {
-	return p->pi_blocked_on ? p->pi_blocked_on->lock : NULL;
+	return rt_mutex_real_waiter(p->pi_blocked_on) ?
+		p->pi_blocked_on->lock : NULL;
 }
 
 /*
@@ -479,7 +485,7 @@ static int rt_mutex_adjust_prio_chain(st
 	 * reached or the state of the chain has changed while we
 	 * dropped the locks.
 	 */
-	if (!waiter)
+	if (!rt_mutex_real_waiter(waiter))
 		goto out_unlock_pi;
 
 	/*
@@ -896,6 +902,23 @@ static int task_blocks_on_rt_mutex(struc
 		return -EDEADLK;
 
 	raw_spin_lock_irqsave(&task->pi_lock, flags);
+
+	/*
+	 * In the case of futex requeue PI, this will be a proxy
+	 * lock. The task will wake unaware that it is enqueueed on
+	 * this lock. Avoid blocking on two locks and corrupting
+	 * pi_blocked_on via the PI_WAKEUP_INPROGRESS
+	 * flag. futex_wait_requeue_pi() sets this when it wakes up
+	 * before requeue (due to a signal or timeout). Do not enqueue
+	 * the task if PI_WAKEUP_INPROGRESS is set.
+	 */
+	if (task != current && task->pi_blocked_on == PI_WAKEUP_INPROGRESS) {
+		raw_spin_unlock_irqrestore(&task->pi_lock, flags);
+		return -EAGAIN;
+	}
+
+	BUG_ON(rt_mutex_real_waiter(task->pi_blocked_on));
+
 	__rt_mutex_adjust_prio(task);
 	waiter->task = task;
 	waiter->lock = lock;
@@ -919,7 +942,7 @@ static int task_blocks_on_rt_mutex(struc
 		rt_mutex_enqueue_pi(owner, waiter);
 
 		__rt_mutex_adjust_prio(owner);
-		if (owner->pi_blocked_on)
+		if (rt_mutex_real_waiter(owner->pi_blocked_on))
 			chain_walk = 1;
 	} else if (rt_mutex_cond_detect_deadlock(waiter, chwalk)) {
 		chain_walk = 1;
@@ -1011,7 +1034,7 @@ static void remove_waiter(struct rt_mute
 {
 	bool is_top_waiter = (waiter == rt_mutex_top_waiter(lock));
 	struct task_struct *owner = rt_mutex_owner(lock);
-	struct rt_mutex *next_lock;
+	struct rt_mutex *next_lock = NULL;
 	unsigned long flags;
 
 	raw_spin_lock_irqsave(&current->pi_lock, flags);
@@ -1036,7 +1059,8 @@ static void remove_waiter(struct rt_mute
 	__rt_mutex_adjust_prio(owner);
 
 	/* Store the lock on which owner is blocked or NULL */
-	next_lock = task_blocked_on_lock(owner);
+	if (rt_mutex_real_waiter(owner->pi_blocked_on))
+		next_lock = task_blocked_on_lock(owner);
 
 	raw_spin_unlock_irqrestore(&owner->pi_lock, flags);
 
@@ -1072,7 +1096,7 @@ void rt_mutex_adjust_pi(struct task_stru
 	raw_spin_lock_irqsave(&task->pi_lock, flags);
 
 	waiter = task->pi_blocked_on;
-	if (!waiter || (waiter->prio == task->prio &&
+	if (!rt_mutex_real_waiter(waiter) || (waiter->prio == task->prio &&
 			!dl_prio(task->prio))) {
 		raw_spin_unlock_irqrestore(&task->pi_lock, flags);
 		return;
--- a/kernel/locking/rtmutex_common.h
+++ b/kernel/locking/rtmutex_common.h
@@ -119,6 +119,8 @@ enum rtmutex_chainwalk {
 /*
  * PI-futex support (proxy locking functions, etc.):
  */
+#define PI_WAKEUP_INPROGRESS	((struct rt_mutex_waiter *) 1)
+
 extern struct task_struct *rt_mutex_next_owner(struct rt_mutex *lock);
 extern void rt_mutex_init_proxy_locked(struct rt_mutex *lock,
 				       struct task_struct *proxy_owner);